OpenAI Launches Open Network Protocol MRC in Collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA
OpenAI announced the launch of a new open network protocol, MRC (Multipath Reliable Connection), in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA.
This protocol is designed for large-scale AI training clusters, significantly improving operational speed and stability while effectively reducing GPU resource waste. It is now fully deployed across all of OpenAI's cutting-edge model training supercomputers, including the joint facility with Oracle in Abilene, Texas, and Microsoft's Fairwater supercomputer.
MRC has been made available to the entire industry through the Open Compute Project for free use and collaborative development.
Source: Public Information
ABAB AI Insight
OpenAI's collaboration with five major chip and cloud giants to launch the MRC protocol continues its strategy of penetrating from model training to underlying infrastructure. Previously, it secured hardware through substantial computing power contracts with Microsoft and Oracle. This open network layer protocol marks its first initiative to lead cross-vendor interconnect standards, aiming to address common network bottlenecks and low GPU utilization in large-scale clusters.
In terms of capital pathways, OpenAI directly translates network optimization of training clusters into overall computing efficiency improvements. By reducing inter-node communication overhead and fault recovery time with MRC, combined with its multi-billion dollar computing power prepayment contracts, the goal is to maximize effective FLOPS per GPU while establishing the protocol as a de facto standard in the industry, thereby consolidating its influence in training infrastructure.
The current AI training cluster is at a critical stage of transitioning from single-vendor networks to multi-vendor open reliable interconnect protocols, similar to the closed evolution of NVIDIA NVLink, Broadcom Tomahawk switch ecosystems, and internal optimizations of Google Jupiter networks.
Essentially, this represents a restructuring of the industry chain: the MRC protocol reconstructs dispersed GPU cluster networks into a unified, efficient layer through multipath reliable connections, shifting capital from hardware procurement to network utilization optimization. Mechanically, it reduces the complexity and fault costs of cluster deployment through open standards, accelerating the structural shift of AI training from single-vendor lock-in to multi-supplier collaboration, providing foundational support for the era of tens of thousands or even hundreds of thousands of GPUs.
ABAB News · Cognitive Law
The true competition in computing power ultimately lies in who can ensure that no GPU is wasted. Open protocols often secure industry dominance more effectively than closed hardware. When training clusters shift from 'buying hardware' to 'building networks', the real competition in infrastructure begins.